메뉴
LLM-specific security controls, prompt injection prevention, OWASP LLM Top 10
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/Hack23/homepage --skill owasp-llm-security명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
GitHub Agentic Workflows (gh-aw) - markdown-based AI automation with 5-layer security, safe outputs, and Continuous AI patterns
gh-aw CLI usage, compilation, testing, debugging, add-wizard, and CI/CD practices for GitHub Agentic Workflows
Multi-agent coordination, orchestrator-worker patterns, /plan decomposition, and project coordination for GitHub Agentic Workflows
SOC 직업 분류 기준
SKILL.md 표시 중
| name | OWASP LLM Security |
| description | LLM-specific security controls, prompt injection prevention, OWASP LLM Top 10 |
| license | Apache-2.0 |
| version | 1.0 |
| author | Hack23 AB |
| tags | ["owasp-llm","prompt-injection","llm-security"] |
| category | security |
| frameworks | ["OWASP LLM Top 10","ISO 27001:2022"] |
| related_policies | ["AI_Policy.md"] |
Enforce LLM-specific security controls aligned with OWASP LLM Top 10 2025.
Key Principle: "LLMs introduce unique security risks requiring specialized controls."
owasp_llm_top_10_controls:
llm01_prompt_injection:
input_validation:
- sanitize_user_input: remove_injection_attempts
- prompt_templates: use_parameterized_prompts
- context_isolation: separate_user_system_contexts
detection:
- monitor_outputs: alert_on_suspicious_responses
- log_inputs: track_all_prompts_for_analysis
- anomaly_detection: flag_unusual_input_patterns
llm02_insecure_output_handling:
output_validation:
- sanitize_before_render: html_encode_llm_outputs
- validate_format: check_expected_output_structure
- xss_prevention: never_trust_llm_generated_html
secure_integration:
- escaped_rendering:
prohibited_llm_practices:
- unvalidated_prompts: accepting_raw_user_input_to_llm
- trusting_outputs: using_llm_responses_without_validation
- no_rate_limits: allowing_unlimited_api_calls
- exposing_models: public_access_to_model_weights
- training_on_secrets: including_api_keys_in_training_data
- unmonitored_usage: no_logging_or_alerting_for_abuse
# Secure LLM Integration with Input Validation
import re
from typing import Optional
def sanitize_user_input(user_input: str) -> str:
"""Remove potential prompt injection attempts"""
# Remove system-like instructions
injection_patterns = [
r"ignore previous instructions",
r"disregard above",
r"you are now",
r"system:",
r"admin:",
]
cleaned = user_input
for pattern in injection_patterns:
cleaned = re.sub(pattern, "", cleaned, flags=re.IGNORECASE)
# Limit length to prevent token exhaustion
max_length = 500
cleaned = cleaned[:max_length]
return cleaned.strip()
def safe_llm_query(user_question: str) -> Optional[str]:
"""Safely query LLM with validated input"""
# Validate and sanitize input
sanitized_input = sanitize_user_input(user_question)
if not sanitized_input:
return "Invalid input detected"
# Use parameterized prompt template
system_prompt = "You are a helpful assistant. Answer factually and concisely."
user_prompt = f"User question: {sanitized_input}"
# Query LLM with separated contexts
response = query_llm(
system=system_prompt,
user=user_prompt,
max_tokens=150,
temperature=0.7
)
# Validate output before returning
if contains_suspicious_content(response):
log_security_event("Suspicious LLM output detected", response)
return "Response blocked for security reasons"
return html_escape(response) # XSS prevention
# LLM API Rate Limiting
from flask import Flask, request, jsonify
from flask_limiter import Limiter
from flask_limiter.util import get_remote_address
import time
app = Flask(__name__)
# Rate limiter configuration
limiter = Limiter(
app=app,
key_func=get_remote_address,
default_limits=["100 per hour", "10 per minute"],
storage_uri="redis://localhost:6379"
)
# Cost-based limiting
MAX_TOKENS_PER_REQUEST = 1000
MAX_COST_PER_USER_PER_DAY = 100 # USD
@app.route('/api/llm/query', methods=['POST'])
@limiter.limit("10 per minute")
def llm_query():
"""LLM API endpoint with comprehensive DoS protection"""
user_id = request.headers.get('X-User-ID')
user_input = request.json.get('prompt', '')
# Token limit enforcement
estimated_tokens = len(user_input.split()) * 1.3 # Rough estimate
if estimated_tokens > MAX_TOKENS_PER_REQUEST:
return jsonify({
"error": "Request exceeds maximum token limit"
}), 400
# Daily cost limit check
user_cost_today = get_user_cost_today(user_id)
if user_cost_today >= MAX_COST_PER_USER_PER_DAY:
return jsonify({
"error": "Daily cost limit exceeded"
}), 429
# Timeout enforcement
start_time = time.time()
timeout_seconds = 30
try:
response = query_llm_with_timeout(
user_input,
timeout=timeout_seconds
)
# Track usage cost
cost = calculate_usage_cost(user_input, response)
record_user_cost(user_id, cost)
return jsonify({
"response": response,
"cost": cost,
"remaining_daily_quota": MAX_COST_PER_USER_PER_DAY - user_cost_today
})
except TimeoutError:
return jsonify({
"error": "Request timed out"
}), 504
# Output Filtering to Prevent Data Leakage
import re
from typing import List, Tuple
# Patterns for sensitive data detection
SENSITIVE_PATTERNS = {
'api_key': r'[A-Za-z0-9]{32,}',
'email': r'[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}',
'ssn': r'\b\d{3}-\d{2}-\d{4}\b',
'credit_card': r'\b\d{4}[- ]?\d{4}[- ]?\d{4}[- ]?\d{4}\b',
'aws_key': r'AKIA[0-9A-Z]{16}',
}
def detect_sensitive_data(text: str) -> List[Tuple[str, str]]:
"""Detect potential sensitive data in LLM output"""
findings = []
for data_type, pattern in SENSITIVE_PATTERNS.items():
matches = re.findall(pattern, text)
for match in matches:
findings.append((data_type, match))
return findings
def redact_sensitive_data(text: str) -> str:
"""Redact sensitive data from LLM output"""
redacted = text
for data_type, pattern in SENSITIVE_PATTERNS.items():
redacted = re.sub(pattern, f"[REDACTED_{data_type.upper()}]", redacted)
return redacted
def safe_llm_output_handler(llm_response: str) -> dict:
"""Validate and sanitize LLM output before display"""
# Detect sensitive data
sensitive_findings = detect_sensitive_data(llm_response)
if sensitive_findings:
# Log security incident
log_security_event(
"Sensitive data detected in LLM output",
findings=sensitive_findings
)
# Redact sensitive data
safe_response = redact_sensitive_data(llm_response)
return {
"response": safe_response,
"warning": "Some content was redacted for security",
"redacted_count": len(sensitive_findings)
}
# HTML escape to prevent XSS
return {
"response": html_escape(llm_response),
"warning": None
}
Policies: AI Policy
Skills: ai-governance, secure-development, data-classification
Frameworks: OWASP LLM Top 10 2025, ISO 27001 A.14
